ComfyUI Semantic Infinite Zoom Workflow

aurelm · reddit · 2026-07-13

The author shares a "semantic infinite zoom" workflow for ComfyUI, aiming not just to upscale an image, but to maintain semantic continuity of the subject across every zoom level.

Core Concept

Before each crop, a Vision Language Model (VLM) is used alongside the original image and the crop to reinterpret what the current region is.
The author uses Qwen VLM because Krea 2 already loads it as a visual encoder in VRAM, making it convenient to reuse for context analysis.
The crop is first upscaled using a traditional GAN upscaler, then regenerated by Krea 2 with very low denoising (around 0.05–0.15) to synthesize finer structural details.
Finally, Flux Kontext Klein is used for cleanup: denoising, sharpening, reducing chromatic aberration, and enhancing local contrast without altering the underlying structure.

Why It Works

The author argues that standard recursive upscaling fails not because of resolution limits, but due to semantic loss: as you zoom deeper, the model forgets what it is looking at and eventually turns an eye into generic texture.

By continuously injecting semantic descriptions like "this is still the same eye of the same eagle" at each step, this workflow ensures detail generation remains anchored to a coherent biological structure.

Final Output

He describes this as "infinite semantic detail"—not infinite pixels, but allowing details like feathers, fibers, keratin, and iris tissue to unfold semantically and continuously.

Original post →

ComfyUI Semantic Infinite Zoom Workflow

More from coding & agent